vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Refactor OpenCL code to work more like the CUDA code, add missing functions

Browse source
This commit is contained in:
0cc4m 2023-05-14 17:01:46 +02:00
parent a7e3bee4cc
commit 17e53dbb7e
6 changed files with 656 additions and 180 deletions
Download patch file Download diff file Expand all files Collapse all files

5
Makefile
View file

@ -138,6 +138,7 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
endif
ifdef LLAMA_CLBLAST
CFLAGS += -DGGML_USE_CLBLAST
CXXFLAGS += -DGGML_USE_CLBLAST
# Mac provides OpenCL as a framework
ifeq ($(UNAME_S),Darwin)
LDFLAGS += -lclblast -framework OpenCL
@ -145,8 +146,8 @@ ifdef LLAMA_CLBLAST
LDFLAGS += -lclblast -lOpenCL
endif
OBJS += ggml-opencl.o
ggml-opencl.o: ggml-opencl.c ggml-opencl.h
$(CC) $(CFLAGS) -c $< -o $@
ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
$(CXX) $(CXXFLAGS) -c $< -o $@
endif
ifneq ($(filter aarch64%,$(UNAME_M)),)
# Apple M1, M2, etc.

697
ggml-opencl.cpp
View file

@ -1,7 +1,9 @@
#include "ggml-opencl.h"
#include <atomic>
#define CL_TARGET_OPENCL_VERSION 110
#include <clblast_c.h>
#include <clblast.h>
#include <stdlib.h>
#include <stdio.h>
@ -9,6 +11,8 @@
#include "ggml.h"
#define CL_DMMV_BLOCK_SIZE 32;
#define MULTILINE_QUOTE(...) #__VA_ARGS__
static const char * program_source = MULTILINE_QUOTE(
@ -52,6 +56,13 @@ struct __attribute__ ((packed)) block_q8_0
};
__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
const uint i = get_global_id(0);
y[i] = vload_half(0, &x[i]);
}
__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) {
const uint i = get_global_id(0) / 32; /* QK4_0 */
const uint j = get_local_id(0);
@ -124,6 +135,53 @@ __kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float*
y[i*32 + j] = x[i].qs[j]*d;
}
__kernel void dequantize_mul_mat_vec(__global struct block_q4_0* x, __local float* tmp, __global float* y, __global float* dst, int ncols) {
const int row = get_global_id(0);
const int tid = get_local_id(0);
const int block_size = get_local_size(0);
const uint qk = 32; /* QK4_0 */
const uint qr = 2; /* QR4_0 */
const int y_offset = qr == 1 ? 1 : qk/2;
tmp[tid] = 0;
for (int i = 0; i < ncols/block_size; i += 2) {
const int col = i*block_size + 2*tid;
const int ib = (row*ncols + col)/qk; // block index
const int iqs = (col%qk)/qr; // quant index
const int iybs = col - col%qk; // y block start index
// dequantize
float v0, v1;
const float d = vload_half(0, &x[ib].d);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
v0 = (vi0 - 8)*d;
v1 = (vi1 - 8)*d;
// matrix multiplication
tmp[tid] += v0 * y[iybs + iqs + 0];
tmp[tid] += v1 * y[iybs + iqs + y_offset];
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=block_size/2; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
#define CL_CHECK(err) \
@ -151,14 +209,16 @@ static cl_device_id device;
static cl_context context;
static cl_command_queue queue;
static cl_program program;
static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q5_0, kernel_q5_1, kernel_q8_0;
static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
static cl_kernel convert_fp16_to_fp32_cl;
static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
static cl_kernel dequantize_mul_mat_vec_cl;
static bool fp16_support;
static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
cl_program p;
char *program_log;
size_t program_size, log_size;
size_t program_size;
size_t log_size;
int err;
program_size = strlen(program_buffer);
@ -185,7 +245,7 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
}
void ggml_cl_init(void) {
cl_int err = 0;
cl_int err;
struct cl_device;
struct cl_platform {
@ -335,6 +395,20 @@ void ggml_cl_init(void) {
platform = default_device->platform->id;
device = default_device->id;
size_t ext_str_size;
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
char* ext_buffer = (char*) malloc(sizeof(char) * ext_str_size);
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
// Check if ext_buffer contains cl_khr_fp16
for (size_t i = 0; i < ext_str_size - 12; i++) {
if (memcmp(ext_buffer + i, "cl_khr_fp16", 11) == 0) {
fp16_support = true;
break;
}
}
free(ext_buffer);
fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
cl_context_properties properties[] = {
(intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
};
@ -348,127 +422,512 @@ void ggml_cl_init(void) {
program = build_program_from_source(context, device, program_source);
// Prepare dequantize kernels
CL_CHECK((kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
CL_CHECK((kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
CL_CHECK((kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
CL_CHECK((kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
CL_CHECK((kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
// FP16 to FP32 kernel
CL_CHECK((convert_fp16_to_fp32_cl = clCreateKernel(program, "convert_fp16_to_fp32", &err), err));
// Dequantize kernels
CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
// dequant mul mat kernel
CL_CHECK((dequantize_mul_mat_vec_cl = clCreateKernel(program, "dequantize_mul_mat_vec", &err), err));
}
static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
if (req_size <= *cur_size) {
return;
static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_row_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_row_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_row_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_row_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_row_q8_0_cl;
case GGML_TYPE_F16:
return &convert_fp16_to_fp32_cl;
default:
return nullptr;
}
}
// Reallocate buffer with enough space
if (*cur_size > 0) {
clReleaseMemObject(*buf);
static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_mul_mat_vec_cl;
// case GGML_TYPE_Q4_1:
// return dequantize_mul_mat_vec_q4_1_cl;
// case GGML_TYPE_Q5_0:
// return dequantize_mul_mat_vec_q5_0_cl;
// case GGML_TYPE_Q5_1:
// return dequantize_mul_mat_vec_q5_1_cl;
// case GGML_TYPE_Q8_0:
// return dequantize_mul_mat_vec_q8_0_cl;
// case GGML_TYPE_F16:
// return convert_mul_mat_vec_f16_cl;
default:
return nullptr;
}
}
// buffer pool for cl
#define MAX_CL_BUFFERS 256
struct scoped_spin_lock {
std::atomic_flag& lock;
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
while (lock.test_and_set(std::memory_order_acquire)) {
; // spin
}
}
~scoped_spin_lock() {
lock.clear(std::memory_order_release);
}
scoped_spin_lock(const scoped_spin_lock&) = delete;
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};
struct cl_buffer {
cl_mem mem;
size_t size = 0;
};
static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size, cl_mem_flags flags) {
scoped_spin_lock lock(g_cl_pool_lock);
cl_int err;
CL_CHECK((*buf = clCreateBuffer(context, flags, req_size, NULL, &err), err));
*cur_size = req_size;
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer& b = g_cl_buffer_pool[i];
if (b.size > 0 && b.size >= size) {
cl_mem mem = b.mem;
*actual_size = b.size;
b.size = 0;
return mem;
}
}
cl_mem mem;
CL_CHECK((mem = clCreateBuffer(context, flags, size, NULL, &err), err));
*actual_size = size;
return mem;
}
void ggml_cl_sgemm_wrapper(
const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
const int m, const int n, const int k,
const float alpha, const void *host_a, const int lda,
const float *host_b, const int ldb, const float beta,
float *host_c, const int ldc, const int btype) {
static void ggml_cl_pool_free(cl_mem mem, size_t size) {
scoped_spin_lock lock(g_cl_pool_lock);
cl_kernel kernel;
size_t global = n * k, local, size_qb;
bool dequant;
switch (btype) {
case GGML_TYPE_F32:
dequant = false;
break;
case GGML_TYPE_Q4_0:
dequant = true;
kernel = kernel_q4_0;
local = 16;
size_qb = global * (sizeof(ggml_fp16_t) + local) / 32;
break;
case GGML_TYPE_Q4_1:
dequant = true;
kernel = kernel_q4_1;
local = 16;
size_qb = global * (sizeof(ggml_fp16_t) * 2 + local) / 32;
break;
case GGML_TYPE_Q5_0:
dequant = true;
kernel = kernel_q5_0;
local = 16;
size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32;
break;
case GGML_TYPE_Q5_1:
dequant = true;
kernel = kernel_q5_1;
local = 16;
size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
break;
case GGML_TYPE_Q8_0:
dequant = true;
kernel = kernel_q8_0;
local = 32;
size_qb = global * (sizeof(ggml_fp16_t) + local) / 32;
break;
default:
fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
abort();
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer& b = g_cl_buffer_pool[i];
if (b.size == 0) {
b.mem = mem;
b.size = size;
return;
}
}
const size_t size_a = m * k * sizeof(float);
const size_t size_b = n * k * sizeof(float);
const size_t size_c = m * n * sizeof(float);
// Prepare buffers
ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
if (dequant) {
ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
}
ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
cl_event ev_a, ev_qb, ev_b;
if (dequant) {
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b));
CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb));
} else {
CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b));
}
CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a));
if (dequant) {
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b));
CL_CHECK(clReleaseEvent(ev_qb));
}
CL_CHECK(clWaitForEvents(1, &ev_a));
CL_CHECK(clWaitForEvents(1, &ev_b));
CL_CHECK(clReleaseEvent(ev_a));
CL_CHECK(clReleaseEvent(ev_b));
cl_event ev_sgemm;
CLBLAST_CHECK(CLBlastSgemm(
(CLBlastLayout)order,
(CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
m, n, k,
alpha,
cl_buffer_a, 0, lda,
cl_buffer_b, 0, ldb,
beta,
cl_buffer_c, 0, ldc,
&queue, &ev_sgemm));
cl_event ev_c;
CL_CHECK(clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c));
// Wait for completion
CL_CHECK(clWaitForEvents(1, &ev_c));
CL_CHECK(clReleaseEvent(ev_sgemm));
CL_CHECK(clReleaseEvent(ev_c));
fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
clReleaseMemObject(mem);
}
static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
cl_int err;
const uint64_t ne0 = src->ne[0];
const uint64_t ne1 = src->ne[1];
const uint64_t nb0 = src->nb[0];
const uint64_t nb1 = src->nb[1];
const uint64_t nb2 = src->nb[2];
const uint64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const size_t ts = ggml_type_size(type);
const size_t bs = ggml_blck_size(type);
const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
if (nb0 == ts && nb1 == ts*ne0/bs) {
err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
return err;
}
if (nb0 == ts) {
const size_t buffer_origin[3] = { offset, 0, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts*ne0/bs, ne1, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
return err;
}
for (uint64_t i1 = 0; i1 < ne1; i1++) {
// pretend the row is a matrix with cols=1
const size_t buffer_origin[3] = { offset, i1, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts/bs, ne0, 1 };
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
if (err != CL_SUCCESS) {
break;
}
}
return err;
}
static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size, CL_MEM_READ_ONLY);
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size, CL_MEM_READ_ONLY);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy data to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
GGML_ASSERT(fp16_support);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
const int nb12 = src1->nb[2];
const int nb13 = src1->nb[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size, CL_MEM_READ_ONLY);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
}
}
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
for (int64_t i00 = 0; i00 < ne10; i00++) {
// very slow due to no inlining
tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
}
}
}
// copy src1 to device
CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host, then convert to float
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
size_t x_size;
size_t y_size;
size_t d_size;
size_t q_size;
cl_mem d_X;
if (!mul_mat_vec) {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size, CL_MEM_READ_WRITE);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size, CL_MEM_READ_ONLY);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
cl_mem d_Q;
if (src0->backend == GGML_BACKEND_CPU) {
d_Q = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
}
cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
GGML_ASSERT(to_fp32_cl != nullptr);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
cl_event ev_Q;
cl_event ev_sgemm;
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, &ev_Q));
} else if (src0->backend == GGML_BACKEND_CL) {
d_Q = * (cl_mem *) src0->data;
} else {
GGML_ASSERT(false);
}
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
printf("Gogogo\n");
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
// compute
// dequantize_mul_mat_vec(__global void * vx, __local float* tmp, __global float * y, __global float * dst, __global int ncols, __global int vx_type) {
const size_t global = ne00;
const size_t local = CL_DMMV_BLOCK_SIZE;
const cl_int ncols = ne01;
const cl_int qtype = src0->type;
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 1, sizeof(float) * local, NULL));
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 3, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 4, sizeof(cl_int), &ncols));
CL_CHECK(clSetKernelArg(dequantize_mul_mat_vec_cl, 5, sizeof(cl_int), &qtype));
CL_CHECK(clEnqueueNDRangeKernel(queue, dequantize_mul_mat_vec_cl, 1, NULL, &global, &local, 1, &ev_Q, &ev_sgemm));
} else { // general dequantization kernel + CLBlast matrix matrix multiplication
// convert src0 to fp32 on device
const size_t global = x_ne;
const size_t local = 16;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, &local, 1, &ev_Q, NULL));
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
// wait for conversion
CL_CHECK(clFinish(queue));
// compute
clblast::StatusCode status = clblast::Gemm(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
if (!mul_mat_vec) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src0->backend == GGML_BACKEND_CPU) {
ggml_cl_pool_free(d_Q, q_size);
}
}
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CL)) {
return true;
}
return false;
}
bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
// If device doesn't support FP16
if (!fp16_support) {
return false;
}
size_t src0_sz = ggml_nbytes(src0);
size_t src1_sz = ggml_nbytes(src1);
// mul_mat_q: src0 is converted to fp32 on device
size_t mul_mat_q_transfer = src0_sz + src1_sz;
// mul_mat_f16: src1 is converted to fp16 on cpu
size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
// choose the smaller one to transfer to the device
// TODO: this is not always the best choice due to the overhead of converting to fp16
return mul_mat_f16_transfer < mul_mat_q_transfer;
}
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) {
ggml_cl_mul_mat_f32(src0, src1, dst);
}
else if (src0->type == GGML_TYPE_F16) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
}
else {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
}
else if (ggml_is_quantized(src0->type)) {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
else {
GGML_ASSERT(false);
}
}
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return ggml_nelements(src1) * sizeof(ggml_fp16_t);
}
return 0;
}
void ggml_cl_transform_tensor(ggml_tensor * tensor) {
const int64_t ne0 = tensor->ne[0];
const int64_t ne1 = tensor->ne[1];
const int64_t ne2 = tensor->ne[2];
const int64_t ne3 = tensor->ne[3];
const ggml_type type = tensor->type;
const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
size_t q_size;
cl_mem* d_Q = (cl_mem*) malloc(sizeof(cl_mem));
*d_Q = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
// copy tensor to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, *d_Q, 0, tensor, 0, 0, NULL));
CL_CHECK(clFinish(queue));
tensor->data = d_Q;
tensor->backend = GGML_BACKEND_CL;
}

18
ggml-opencl.h
View file

@ -1,23 +1,21 @@
#pragma once
#include "ggml.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_cl_init(void);
enum ggml_blas_order {
GGML_BLAS_ORDER_ROW_MAJOR = 101,
GGML_BLAS_ORDER_COLUMN_MAJOR = 102,
};
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
enum ggml_blas_op {
GGML_BLAS_OP_N = 111,
GGML_BLAS_OP_T = 112,
GGML_BLAS_OP_C = 113,
};
void * ggml_cl_host_malloc(size_t size);
void ggml_cl_host_free(void * ptr);
void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
#ifdef __cplusplus
}

83
ggml.c
View file

@ -9431,7 +9431,7 @@ static void ggml_compute_forward_rms_norm_back(
// ggml_compute_forward_mul_mat
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
// helper function to determine if it is better to use BLAS or not
// for large matrices, BLAS is faster
static bool ggml_compute_forward_mul_mat_use_blas(
@ -9472,7 +9472,7 @@ static void ggml_compute_forward_mul_mat_f32(
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
const int64_t ne10 = src1->ne[0];
#endif
const int64_t ne11 = src1->ne[1];
@ -9536,9 +9536,16 @@ static void ggml_compute_forward_mul_mat_f32(
}
return;
}
#elif defined(GGML_USE_CLBLAST)
if (ggml_cl_can_mul_mat(src0, src1, dst)) {
if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
}
return;
}
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
if (params->ith != 0) {
return;
@ -9558,21 +9565,11 @@ static void ggml_compute_forward_mul_mat_f32(
const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
#if defined(GGML_USE_CLBLAST)
// zT = y * xT
ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
ne11, ne01, ne10,
1.0f, y, ne10,
x, ne10,
0.0f, d, ne01,
GGML_TYPE_F32);
#else
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
ne11, ne01, ne10,
1.0f, y, ne10,
x, ne00,
0.0f, d, ne01);
#endif
}
}
//printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
@ -9711,9 +9708,16 @@ static void ggml_compute_forward_mul_mat_f16_f32(
}
return;
}
#elif defined(GGML_USE_CLBLAST)
if (ggml_cl_can_mul_mat(src0, src1, dst)) {
if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
}
return;
}
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
GGML_ASSERT(nb10 == sizeof(float));
@ -9743,20 +9747,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
assert(id*sizeof(float) <= params->wsize);
}
#if defined(GGML_USE_CLBLAST)
const float * x = wdata;
const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
// zT = y * xT
ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
ne11, ne01, ne10,
1.0f, y, ne10,
x, ne10,
0.0f, d, ne01,
GGML_TYPE_F32);
#else
const float * x = wdata;
const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
@ -9768,7 +9758,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
1.0f, y, ne10,
x, ne00,
0.0f, d, ne01);
#endif
}
}
@ -9931,9 +9920,16 @@ static void ggml_compute_forward_mul_mat_q_f32(
}
return;
}
#elif defined(GGML_USE_CLBLAST)
if (ggml_cl_can_mul_mat(src0, src1, dst)) {
if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
}
return;
}
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
if (params->ith != 0) {
return;
@ -9956,9 +9952,6 @@ static void ggml_compute_forward_mul_mat_q_f32(
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
#if defined(GGML_USE_CLBLAST)
const void* x = (char *) src0->data + i03*nb03 + i02*nb02;
#else
{
size_t id = 0;
for (int64_t i01 = 0; i01 < ne01; ++i01) {
@ -9970,23 +9963,12 @@ static void ggml_compute_forward_mul_mat_q_f32(
}
const float * x = wdata;
#endif
#if defined(GGML_USE_CLBLAST)
// zT = y * xT
ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
ne11, ne01, ne10,
1.0f, y, ne10,
x, ne10,
0.0f, d, ne01,
type);
#else
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
ne11, ne01, ne10,
1.0f, y, ne10,
x, ne00,
0.0f, d, ne01);
#endif
}
}
@ -14165,9 +14147,16 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
cur = ggml_cuda_mul_mat_get_wsize(node->src0, node->src1, node);
}
else
#elif defined(GGML_USE_CLBLAST)
if (ggml_cl_can_mul_mat(node->src0, node->src1, node)) {
node->n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
cur = ggml_cl_mul_mat_get_wsize(node->src0, node->src1, node);
}
else
#endif
if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) {
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
@ -14181,13 +14170,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
#endif
} else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
cur = 0;
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1;
}
#endif
} else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) {
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1;
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);

1
ggml.h
View file

@ -249,6 +249,7 @@ extern "C" {
enum ggml_backend {
GGML_BACKEND_CPU = 0,
GGML_BACKEND_CUDA = 1,
GGML_BACKEND_CL = 2,
};
// model file types

32
llama.cpp
View file

@ -12,6 +12,8 @@
#include "ggml.h"
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h"
#endif
#include <array>
@ -1092,7 +1094,7 @@ static void llama_model_load_internal(
fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
}
fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
#else
#elif !defined(GGML_USE_CLBLAST)
(void) n_gpu_layers;
#endif
}
@ -1125,7 +1127,33 @@ static void llama_model_load_internal(
done_size += lt.size;
}
}
#endif // GGML_USE_CUBLAS
#elif defined(GGML_USE_CLBLAST)
{
const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
fprintf(stderr, "ggml_opencl: offloading %d layers to GPU\n", n_gpu);
size_t vram_total = 0;
for (int i = 0; i < n_gpu; ++i) {
const auto & layer = model.layers[i];
ggml_cl_transform_tensor(layer.wq); vram_total += ggml_nbytes(layer.wq);
ggml_cl_transform_tensor(layer.wk); vram_total += ggml_nbytes(layer.wk);
ggml_cl_transform_tensor(layer.wv); vram_total += ggml_nbytes(layer.wv);
ggml_cl_transform_tensor(layer.wo); vram_total += ggml_nbytes(layer.wo);
ggml_cl_transform_tensor(layer.w1); vram_total += ggml_nbytes(layer.w1);
ggml_cl_transform_tensor(layer.w2); vram_total += ggml_nbytes(layer.w2);
ggml_cl_transform_tensor(layer.w3); vram_total += ggml_nbytes(layer.w3);
}
if (n_gpu_layers > (int) hparams.n_layer) {
fprintf(stderr, "ggml_opencl: offloading output layer to GPU\n");
ggml_cl_transform_tensor(model.output); vram_total += ggml_nbytes(model.output);
}
fprintf(stderr, "ggml_opencl: total VRAM used: %zu MB\n", vram_total / 1024 / 1024);
}
#endif
if (progress_callback) {
progress_callback(1.0f, progress_callback_user_data);